[ExI] More super forecasting

[John Clark]

*In the name of completeness I figured I should ask Gemini the same
question that I asked Claude and GPT, you can see its Full response below
but Gemini's  last line was: "Forecast: 18% probability of a positive
resolution by December 31, 2030".*

*ME:**I want you to think like a super forecaster when answering this
question. "By December 31, 2030, does there exist an AI system, possibly
using external tools, code execution, retrieval, and orchestration of other
AI models but not specialized robotic hardware, that is recognized by broad
expert and scientific consensus as matching or exceeding the best
documented human performance, within roughly comparable time and cost, on
essentially every cognitive task that has an objective or
intersubjectively-agreed success criterion (math, science, programming,
forecasting, strategy games, engineering, legal/medical/financial analysis,
etc.) — doing so at least as consistently as elite human experts — with
tasks gated by legal or biological status, tasks with no objective quality
standard, and tasks bottlenecked by physical embodiment (not just fine
dexterity, but also raw speed, strength, and endurance) all excluded?*

*Gemini: "*To approach this question with the rigor of a superforecaster,
we must strip away both techno-optimist hype and biological exceptionalism.
We must focus strictly on the resolution criteria, weigh the "inside view"
(technological trajectory) against the "outside view" (historical base
rates of technological consensus), and identify the critical bottlenecks.

Here is a probabilistic breakdown of why this specific, highly constrained
definition of AGI (often referred to as "Strong Drop-in AGI") is a
formidable hurdle by December 31, 2030.
. Deconstructing the Resolution Criteria

The prompt establishes an incredibly high bar. To resolve positively, the
system must not just be "smart"; it must meet all the following conditions
simultaneously:

   -

   *Peak Capability:* Matches or exceeds the *best documented human
   performance* (not just the average professional, but the elite).
   -

   *Universality:* Applies to *essentially every* objective cognitive task
   (math, engineering, strategy, law, finance).
   -

   *Reliability:* Acts at least as consistently as elite humans.
   -

   *Efficiency:* Operates at roughly comparable time and cost.
   -

   *Sociological Threshold:* Recognized by *broad expert and scientific
   consensus*.
   2. The Case for Positive Resolution (The "Inside View")

   There are strong trend lines suggesting massive capability overhangs
   that could be unlocked by 2030 (roughly 4.5 years from today, July 2026).
   -

      *Compute Scaling & Investment:* Trillion-dollar infrastructure
      investments are already in motion. If scaling laws (predictable
performance
      improvements relative to compute/data) hold through the next two
orders of
      magnitude,* the raw pattern-matching and generation capabilities of
      models will be staggering*.
      -

      *System 2 / Test-Time Compute:* We are already seeing the shift from
      pure autoregressive next-token prediction to "reasoning" models that
      utilize reinforcement learning and test-time compute to explore solution
      trees (e.g., AlphaGeometry, early self-play reasoning models).
      -

      *Tool Use and Orchestration:* The prompt explicitly allows external
      tools and code execution. AI does not need to perfectly simulate
a physics
      engine in its weights; it just needs to know how to write the
code to run a
      Python simulation. Agentic scaffolding (AI orchestrating other AI tools)
      drastically lowers the barrier to general capability.
      3. The Cruxes (Why this likely resolves negatively by 2030)

      Despite rapid progress, several severe bottlenecks make a positive
      resolution by 2030 highly unlikely based on this strict definition.
      A. The "Best Documented Human" Benchmark

      Matching a 95th percentile programmer or a junior lawyer is
      fundamentally different from matching the *best documented human
      performance*. The elite human edge in fields like theoretical
      physics, novel mathematics (e.g., Terence Tao), or high-stakes financial
      strategy relies on extreme abstraction, out-of-distribution
intuition, and
      multi-year focused persistence. Current AI architectures are
exceptional at
      interpolating within human data distributions, but generating world-class
      *extrapolative* novel insights consistently is structurally difficult
      for models trained on existing data.
      B. The Error-Compounding Bottleneck in Long-Horizon Tasks Elite human
      tasks are rarely single-turn queries. Engineering a novel aerospace
      component or conducting a multi-month scientific research
project requires
      executing tens of thousands of sequential steps. Even if an AI
agent has a
      99.9% success rate per step, over a 1,000-step autonomous task, the
      probability of successful completion without catastrophic
compounding error
      drops to ~36%. Elite humans self-correct intuitively; AI agents currently
      struggle to recognize when they have drifted fundamentally off track over
      long time horizons.C. The "Comparable Time and Cost" Constraint

      To match an elite researcher, an AI might need to generate millions
      of reasoning trees and simulate countless dead ends. While compute is
      scaling, the *cost* of running massive ensembles of frontier models
      for long-horizon agentic tasks remains exorbitant. By 2030,
achieving peak
      human capability might be technically possible, but doing so within the
      "comparable cost" of hiring an elite human (who runs on ~20 watts of
      biological power) is a massive economic headwind.
      D. The Sociological Bottleneck (Broad Consensus)

      This is perhaps the strongest barrier. Even if the requisite AI
      system is brought online in early 2029, achieving *broad expert and
      scientific consensus* by December 31, 2030, is nearly impossible.
      Scientific consensus requires rigorous peer-reviewed benchmarking,
      replication of results, and exhausting debates over whether the
AI actually
      "solved" the tasks or merely regurgitated contaminated training data.
      Skeptics will continually move the goalposts, demanding multi-year
      longitudinal studies to prove the AI matches elite human consistency.
      Eighteen months is generally not enough time for the scientific community
      to reach consensus on a paradigm-shattering claim.
      4. Synthesis and Forecast
         -

         *Base Rate:* Historically, the final 10% of autonomous reliability
         in complex systems (e.g., self-driving cars) takes vastly
longer than the
         first 90%.
         -

         *Timeline:* We have ~4.5 years from mid-2026. While we will likely
         see AI systems match elite humans in *narrow, highly structured*
         objective tasks (like proving specific mathematical theorems
or dominating
         complex strategy games), generalizing this to *essentially every*
         objective task—especially those requiring long-horizon novel
research—will
         take longer to engineer.
         -

         FactorProbability Adjustment*Compute & Algorithmic Scaling*++
         (Pushes probability up)
         -

         *Tool Use & Agent Orchestration*++ (Pushes probability up)
         -

         *Requirement: "Best documented human"*-- (Pushes probability down)
         -

         *Requirement: "Long-horizon reliability"*--- (Strongly pushes down)
         -

         *Requirement: "Broad scientific consensus"*--- (Strongly pushes
         down)


         -

         *Final Assessment:* While highly capable, economically
         transformative AI agents will undoubtedly exist by 2030, a system that
         meets the absolute upper-bound criteria of this
prompt—matching the *greatest
         living human minds* across *all* objective fields, reliably, at
         cost-parity, with universal scientific agreement—is highly improbable
         within this timeframe.
         -

         *Forecast:* *18% probability of a positive resolution by December
         31, 2030."*


*John K Clark    See what's on my list at  Extropolis
<https://groups.google.com/g/extropolis>*
ycj
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260706/9758aff2/attachment-0001.htm>